|
I believe Brooks will eventually be proven wrong about AI and Automatic Programming. Granted, he has not been proven wrong yet. It is interesting, perhaps even a little surprising, that reasoning about programs has proven to be one of the most intractable challenges for AI. Why, to take a simple-sounding example, when my program does something wrong, can I not simply ask the machine why it did it? Why do I have to go in with a debugger and find the problem myself? This wouldn't require solutions to any of the familiar bugaboos of AI: it doesn't need a massive database of facts about the world, it doesn't require visual image recognition, it doesn't require natural language understanding, it doesn't require robotics or simulated emotions or any of those things. It just requires logical reasoning. Can't computers do that? Well, no, it turns out, they can't do it, not in its full generality. Despite all the things it doesn't require, general logical reasoning is still almost AI-complete, which is a cute way of saying we still don't know how to make machines do it. And fully general reasoning is required to solve the problem. We have lots of static analysis algorithms that can answer specific kinds of questions about programs, but we're still not very close to being able to answer an arbitrary question, even fairly simple ones. The spaces we would have to search are just too big; the branching factors are too high. I think there is hope on the horizon, though. The so-called automated reasoning systems that have been built over the last few decades -- also called theorem provers -- have mostly not made use of any machine learning techniques. These two technologies are starting to be combined, and machine learning is of course a very hot area right now. I believe that eventually we will have Automatic Programming worthy of the name as a result of this combination. |